Method for Training and Testing Obfuscation Network Capable of Obfuscating Data to Protect Personal Information, and Learning Device and Testing Device Using the Same

1. A method for training an obfuscation network capable of obfuscating original data to protect personal information, comprising steps of: (a) inputting, by a learning device, training data into an obfuscation network having one or more previous updated obfuscation parameters, to thereby allow the obfuscation network to obfuscate the training data by using the previous updated obfuscation parameters and thus to generate obfuscated data for training; (b) performing, by the learning device, (i) a process of inputting the obfuscated data for training into a discriminator, capable of determining whether its inputted data is real or fake, to thereby allow the discriminator to output a current obfuscation score for training, representing whether the obfuscated data for training is real or fake, by using one or more current updated determination parameters and (ii) (ii-1) a process of inputting first sub-data for training into a learning network having one or more current updated learning parameters, to thereby allow the learning network to apply a learning operation to the first sub-data for training by using the current updated learning parameters and thus to output first sub characteristic information for training and a process of updating the current updated learning parameters of the learning network to first sub updated learning parameters such that at least one first sub-error, calculated by referring to (1) the first sub characteristic information for training or a first sub task specific output for training created by using the first sub characteristic information for training and (2) a ground truth of the training data, is minimized and (ii-2) while increasing an integer k from 2 to n, where is n is an integer greater than 2, a process of inputting k-th sub-data for training into the learning network having one or more (k−1)-th sub updated learning parameters, to thereby allow the learning network to apply the learning operation to the k-th sub-data for training by using the (k−1)-th sub updated learning parameters and thus to output k-th sub characteristic information for training and a process of updating the (k−1)-th sub updated learning parameters of the learning network to k-th sub updated learning parameters such that at least one k-th sub-error, calculated by referring to (1) the k-th sub characteristic information for training or a k-th sub task specific output for training created by using the k-th sub characteristic information for training and (2) the ground truth of the training data, is minimized, to thereby perform a process of allowing the learning network having the current updated learning parameters to be sub-trained n times, wherein the first sub-data for training to the n-th sub-data for training are selected among the training data and the obfuscated data for training; and (c) updating, by the learning device, the previous updated obfuscation parameters of the obfuscation network to current updated obfuscation parameters such that at least one representative sub-error, calculated by referring to at least part of the first sub-error to the n-th sub-error created in the process of allowing the learning network to be sub-trained n times, is minimized and such that the current obfuscation score for training is maximized.

2. The method of claim 1 , before the step of (b), further comprising a step of: (b-1) performing, by the learning device, (i) (i-1) a process of inputting the training data or the modified data for training into the discriminator having one or more previous updated determination parameters, to thereby allow the discriminator to output a previous modification score for training, representing whether the training data or the modified data for training is real or fake, by using the previous updated determination parameters, wherein the modified data for training is created by modifying the training data or the obfuscated data for training, (i-2) a process of inputting the obfuscated data for training into the discriminator having the previous updated determination parameters, to thereby allow the discriminator to output a previous obfuscation score for training, representing whether the obfuscated data for training is real or fake, by using the previous updated determination parameters, and (i-3) a process of updating the previous updated determination parameters of the discriminator to the current updated determination parameters such that the previous modification score for training is maximized and such that the previous obfuscation score for training is minimized and (ii) a process of inputting the obfuscated data for training into the learning network having one or more previous updated learning parameters, to thereby allow the learning network to apply the learning operation to the obfuscated data for training by using the previous updated learning parameters and thus to output main characteristic information for training and a process of updating the previous updated learning parameters of the learning network to the current updated learning parameters such that at least one main error, calculated by referring to (1) the main characteristic information for training or a main task specific output for training created by using the main characteristic information for training and (2) the ground truth of the training data, is minimized.

3. The method of claim 1 , wherein at least one of the first sub-data for training to the n-th sub-data for training is the training data.

4. The method of claim 1 , wherein the first sub-data for training to the (n−1)-th sub-data for training are the obfuscated data for training and wherein the n-th sub-data for training is the training data.

5. The method of claim 1 , wherein the learning device generates the representative sub-error by summation or averaging of the first sub-error to the n-th sub-error.

6. The method of claim 1 , wherein the learning device generates the representative sub-error by weighted summation of the first sub-error to the n-th sub-error.

7. The method of claim 6 , wherein the learning device allows a weight of one part of sub-errors corresponding to the training data and a weight of a remaining part of sub-errors corresponding to the obfuscated data for training to be different.

8. The method of claim 1 , wherein a maximum of the previous modification score for training is 1 as a value for determining the training data or the modified data for training as real by the discriminator and wherein a minimum of the previous obfuscation score for training is 0 as a value for determining the obfuscated data for training as fake by the discriminator.

9. A method for testing an obfuscation network capable of obfuscating original data to protect personal information, comprising steps of: (a) on condition that a learning device has performed or supported another device to perform (a1) a process of inputting training data into the obfuscation network having the previous updated obfuscation parameters, to thereby allow the obfuscation network to obfuscate the training data by using the previous updated obfuscation parameters and thus to generate obfuscated data for training; (a2) (i) a process of inputting the obfuscated data for training into the discriminator, capable of determining whether its inputted data is real or fake, to thereby allow the discriminator to output a current obfuscation score for training, representing whether the obfuscated data for training is real or fake, by using the current updated determination parameters and (ii) (ii-1) a process of inputting first sub-data for training into the learning network having the current updated learning parameters, to thereby allow the learning network to apply the learning operation to the first sub-data for training by using the current updated learning parameters and thus to output first sub characteristic information for training and a process of updating the current updated learning parameters of the learning network to first sub updated learning parameters such that the first sub-error, calculated by referring to (1) the first sub characteristic information for training or the first sub task specific output for training created by using the first sub characteristic information for training and (2) the ground truth of the training data, is minimized and (ii-2) while increasing the integer k from 2 to n, where n is an integer greater than 2, a process of inputting k-th sub-data for training into the learning network having the (k−1)-th sub updated learning parameters, to thereby allow the learning network to apply the learning operation to the k-th sub-data for training by using the (k−1)-th sub updated learning parameters and thus to output the k-th sub characteristic information for training and a process of updating the (k−1)-th sub updated learning parameters of the learning network to k-th sub updated learning parameters such that at least one k-th sub-error, calculated by referring to (1) the k-th sub characteristic information for training or a k-th sub task specific output for training created by using the k-th sub characteristic information for training and (2) the ground truth of the training data, is minimized, to thereby perform a process of allowing the learning network having the current updated learning parameters to be sub-trained n times, wherein the first sub-data for training to the n-th sub-data for training are selected among the training data and the obfuscated data for training; and (a3) a process of updating the previous updated obfuscation parameters of the obfuscation network to current updated obfuscation parameters such that the representative sub-error, calculated by referring to at least part of the first sub-error to the n-th sub-error created in the process of allowing the learning network to be sub-trained n times, is minimized and such that the current obfuscation score for training is maximized, performing, by a testing device, a process of acquiring test data; and (b) inputting, by the testing device, the test data into the obfuscation network, to thereby allow the obfuscation network to obfuscate the test data by using the learned obfuscation parameters of the obfuscation network and thus to output obfuscated data for testing.

10. The method of claim 9 , wherein, before the step of (a2), the discriminator and the learning network have been trained by the learning device through (i) (i-1) a process of inputting the training data or modified data for training into the discriminator having one or more previous updated determination parameters, to thereby allow the discriminator to output a previous modification score for training, representing whether the training data or the modified data for training is real or fake, by using the previous updated determination parameters, wherein the modified data for training is created by modifying the training data or the obfuscated data for training, (i-2) a process of inputting the obfuscated data for training into the discriminator having the previous updated determination parameters, to thereby allow the discriminator to output a previous obfuscation score for training, representing whether the obfuscated data for training is real or fake, by using the previous updated determination parameters, and (i-3) a process of updating the previous updated determination parameters to the current updated determination parameters such that the previous modification score for training is maximized and such that the previous obfuscation score for training is minimized and (ii) a process of inputting the obfuscated data for training into the learning network having one or more previous updated learning parameters, to thereby allow the learning network to apply the learning operation to the obfuscated data for training by using the previous updated learning parameters and thus to output main characteristic information for training and a process of updating the previous updated learning parameters to the current updated learning parameters such that at least one main error, calculated by referring to (1) the main characteristic information for training or a main task specific output for training created by using the main characteristic information for training and (2) the ground truth of the training data, is minimized.

11. A learning device for training an obfuscation network capable of obfuscating original data to protect personal information, comprising: at least one memory that stores instructions; and at least one processor configured to execute the instructions to perform: (I) a process of inputting training data into an obfuscation network having one or more previous updated obfuscation parameters, to thereby allow the obfuscation network to obfuscate the training data by using the previous updated obfuscation parameters and thus to generate obfuscated data for training, (II) (i) a process of inputting the obfuscated data for training into a discriminator, capable of determining whether its inputted data is real or fake, to thereby allow the discriminator to output a current obfuscation score for training, representing whether the obfuscated data for training is real or fake, by using one or more current updated determination parameters and (ii) (ii-1) a process of inputting first sub-data for training into a learning network having one or more current updated learning parameters, to thereby allow the learning network to apply a learning operation to the first sub-data for training by using the current updated learning parameters and thus to output first sub characteristic information for training and a process of updating the current updated learning parameters of the learning network to first sub updated learning parameters such that at least one first sub-error, calculated by referring to (1) the first sub characteristic information for training or a first sub task specific output for training created by using the first sub characteristic information for training and (2) a ground truth of the training data, is minimized and (ii-2) while increasing an integer k from 2 to n, where n is an integer greater than 2, a process of inputting k-th sub-data for training into the learning network having one or more (k−1)-th sub updated learning parameters, to thereby allow the learning network to apply the learning operation to the k-th sub-data for training by using the (k−1)-th sub updated learning parameters and thus to output k-th sub characteristic information for training and a process of updating the (k−1)-th sub updated learning parameters of the learning network to k-th sub updated learning parameters such that at least one k-th sub-error, calculated by referring to (1) the k-th sub characteristic information for training or a k-th sub task specific output for training created by using the k-th sub characteristic information for training and (2) the ground truth of the training data, is minimized, to thereby perform a process of allowing the learning network having the current updated learning parameters to be sub-trained n times, wherein the first sub-data for training to the n-th sub-data for training are selected among the training data and the obfuscated data for training, and (III) a process of updating the previous updated obfuscation parameters of the obfuscation network to current updated obfuscation parameters such that at least one representative sub-error, calculated by referring to at least part of the first sub-error to the n-th sub-error created in the process of allowing the learning network to be sub-trained n times, is minimized and such that the current obfuscation score for training is maximized.

12. The learning device of claim 11 , wherein, before the process of (II), the processor further performs: (II-1) (i) (i-1) a process of inputting the training data or modified data for training into the discriminator having one or more previous updated determination parameters, to thereby allow the discriminator to output a previous modification score for training, representing whether the training data or the modified data for training is real or fake, by using the previous updated determination parameters, wherein the modified data for training is created by modifying the training data or the obfuscated data for training, (i-2) a process of inputting the obfuscated data for training into the discriminator having the previous updated determination parameters, to thereby allow the discriminator to output a previous obfuscation score for training, representing whether the obfuscated data for training is real or fake, by using the previous updated determination parameters, and (i-3) a process of updating the previous updated determination parameters of the discriminator to the current updated determination parameters such that the previous modification score for training is maximized and such that the previous obfuscation score for training is minimized and (ii) a process of inputting the obfuscated data for training into the learning network having one or more previous updated learning parameters, to thereby allow the learning network to apply the learning operation to the obfuscated data for training by using the previous updated learning parameters and thus to output main characteristic information for training and a process of updating the previous updated learning parameters of the learning network to the current updated learning parameters such that at least one main error, calculated by referring to (1) the main characteristic information for training or a main task specific output for training created by using the main characteristic information for training and (2) the ground truth of the training data, is minimized.

13. The learning device of claim 11 , wherein at least one of the first sub-data for training to the n-th sub-data for training is the training data.

14. The learning device of claim 11 , wherein the first sub-data for training to the (n−1)-th sub-data for training are the obfuscated data for training and wherein the n-th sub-data for training is the training data.

15. The learning device of claim 11 , wherein the processor generates the representative sub-error by summation or averaging of the first sub-error to the n-th sub-error.

16. The learning device of claim 11 , wherein the processor generates the representative sub-error by weighted summation of the first sub-error to the n-th sub-error.

17. The learning device of claim 16 , wherein the processor allows a weight of one part of sub-errors corresponding to the training data and a weight of a remaining part of sub-errors corresponding to the obfuscated data for training to be different.

18. The learning device of claim 11 , wherein a maximum of the previous modification score for training is 1 as a value for determining the training data or the modified data for training as real by the discriminator and wherein a minimum of the previous obfuscation score for training is 0 as a value for determining the obfuscated data for training as fake by the discriminator.

19. A testing device for testing an obfuscation network capable of obfuscating original data to protect personal information, comprising: at least one memory that stores instructions; and at least one processor configured to execute the instructions to perform: (I) on condition that a learning device has performed (I1) a process of inputting training data into the obfuscation network having the previous updated obfuscation parameters, to thereby allow the obfuscation network to obfuscate the training data by using the previous updated obfuscation parameters and thus to generate obfuscated data for training, (I2) (i) a process of inputting the obfuscated data for training into the discriminator, capable of determining whether its inputted data is real or fake, to thereby allow the discriminator to output a current obfuscation score for training, representing whether the obfuscated data for training is real or fake, by using the current updated determination parameters and (ii) (ii-1) a process of inputting first sub-data for training into the learning network having the current updated learning parameters, to thereby allow the learning network to apply the learning operation to the first sub-data for training by using the current updated learning parameters and thus to output first sub characteristic information for training and a process of updating the current updated learning parameters of the learning network to first sub updated learning parameters such that the first sub-error, calculated by referring to (1) the first sub characteristic information for training or the first sub task specific output for training created by using the first sub characteristic information for training and (2) the ground truth of the training data, is minimized and (ii-2) while increasing the integer k from 2 to n, where n is an integer greater than 2, a process of inputting k-th sub-data for training into the learning network having the (k−1)-th sub updated learning parameters, to thereby allow the learning network to apply the learning operation to the k-th sub-data for training by using the (k−1)-th sub updated learning parameters and thus to output the k-th sub characteristic information for training and a process of updating the (k−1)-th sub updated learning parameters of the learning network to k-th sub updated learning parameters such that at least one k-th sub-error, calculated by referring to (1) the k-th sub characteristic information for training or a k-th sub task specific output for training created by using the k-th sub characteristic information for training and (2) the ground truth of the training data, is minimized, to thereby perform a process of allowing the learning network having the current updated learning parameters to be sub-trained n times, wherein the first sub-data for training to the n-th sub-data for training are selected among the training data and the obfuscated data for training, and (I3) a process of updating the previous updated obfuscation parameters of the obfuscation network to current updated obfuscation parameters such that the representative sub-error, calculated by referring to at least part of the first sub-error to the n-th sub-error created in the process of allowing the learning network to be sub-trained n times, is minimized and such that the current obfuscation score for training is maximized, a process of acquiring test data and (II) a process of inputting the test data into the obfuscation network, to thereby allow the obfuscation network to obfuscate the test data by using the learned obfuscation parameters of the obfuscation network and thus to output obfuscated data for testing.

20. The testing device of claim 19 , wherein, before the process of (I2), the discriminator and the learning network have been trained by the learning device through (i) (i-1) a process of inputting the training data or modified data for training into the discriminator having one or more previous updated determination parameters, to thereby allow the discriminator to output a previous modification score for training, representing whether the training data or the modified data for training is real or fake, by using the previous updated determination parameters, wherein the modified data for training is created by modifying the training data or the obfuscated data for training, (i-2) a process of inputting the obfuscated data for training into the discriminator having the previous updated determination parameters, to thereby allow the discriminator to output a previous obfuscation score for training, representing whether the obfuscated data for training is real or fake, by using the previous updated determination parameters, and (i-3) a process of updating the previous updated determination parameters to the current updated determination parameters such that the previous modification score for training is maximized and such that the previous obfuscation score for training is minimized and (ii) a process of inputting the obfuscated data for training into the learning network having one or more previous updated learning parameters, to thereby allow the learning network to apply the learning operation to the obfuscated data for training by using the previous updated learning parameters and thus to output main characteristic information for training and a process of updating the previous updated learning parameters to the current updated learning parameters such that at least one main error, calculated by referring to (1) the main characteristic information for training or a main task specific output for training created by using the main characteristic information for training and (2) the ground truth of the training data, is minimized.